Imagine an AI service that cuts the electricity needed for a fixed task in half. Then imagine that it performs 3 times as many of those tasks. This is a made-up example, not a claim about any company's footprint. The service now uses 50% more electricity than before.
The new process is also saving half the electricity that doing 3 times the work would have used under the old process. Both statements are true. They compare different things, and they can send an argument in opposite directions if only one gets mentioned.
AI's energy debate needs both. Engineers deserve credit for doing useful work with less. Utilities and the public still need to know how much power the growing system will need. A better number per task cannot settle the second question by itself.
Lower use per task can mean higher totals
A claim that a tool uses less power can travel from a research paper into a sales pitch, then into an argument about whether anyone should worry about demand. Somewhere along that journey, the task being measured can fall out of view.
A system may get better at a fixed test while its users ask it to do harder work. Counting each request as one unit does not solve that problem. A short answer and an agent that makes repeated model calls are both requests from someone's point of view. The computer does more work for one than for the other.
In April 2026, the International Energy Agency looked at how much less power AI needs per task. It also looked at growth in use and at new tasks that take more work. Its central forecast puts global data-center electricity use at about 950 terawatt-hours in 2030, up from an estimated 485 in 2025. That forecast covers all data centers, not AI alone, and it is a projection rather than a meter reading from the future.
It gives planners a different kind of information from a result on a fixed test. Even a service that has made large gains may need more power as its use grows. Whether that growth is worth the cost depends on the work being done. The growth cannot be wished away by pointing to a tool that uses less power for yesterday's task.
There is an equally important mistake in the other direction: treating higher total demand as evidence that the engineering accomplished nothing.
In its 2025 study, the IEA modeled a High Efficiency case that met the same demand for digital services with more than 15% less electricity in 2035 than its baseline. That older case models what could happen. It does not claim that power use has already fallen. Demand for services stays the same in both cases, so we can see how much less power the better tools would need.
Those savings can reduce the amount of new supply needed to serve a given workload. For a system facing limits, that is worth doing even while demand grows. Planners still have to allow for that growth when deciding how much supply to build.
Rebound is a claim about cause
The tempting response is to invoke the rebound effect and declare the matter settled. Cheaper computing encourages more computing, so the savings will be eaten. Sometimes the argument goes further: making AI more efficient must eventually increase its energy use.
That certainty runs ahead of the evidence.
Researchers Jonathan Koomey and Eric Masanet have challenged the habit of treating fast growth in use as proof of efficiency-driven rebound. They have good reason to object. A service can attract new users because it has become useful for new jobs, because access has spread or because people are still learning what to do with it. Growth after a gain in efficiency does not tell us how much growth that gain caused.
To say one caused the other, we would need to ask how demand would have changed without the gain. We cannot blame every new request on lower energy costs. Nor can we assume that those lower costs were passed through to users in a form that changed their behavior.
Rebound could still occur, and we should study it. If a service needs less power, costs less to use and gets more use as a result, some of the expected savings can be lost. Losing some savings is different from losing all of them. For total power use to exceed what it would have been without the gain, that extra use has to be large enough to outweigh the saving.
Sasha Luccioni, Emma Strubell and Kate Crawford argue that we need to assess more than direct resource use. We also need to look at what drives companies and people to use AI and at its indirect effects. Taking that concern seriously means finding out what companies do with their gains. We still need evidence to show how much rebound occurs.
A company might use the gains from a cheaper process in several ways:
- Provide the same service with less power.
- Spend the saving on more work for every request.
- Introduce uses that previously cost too much.
Those choices can have different results. A claim that the savings must go to waste tells us little about them.
We still need to track total demand. Even if efficiency causes no extra demand at all, growth for other reasons can still raise power use. We do not have to prove that the saving created the growth before planning for the growth.
Keep efficiency and count the total
A business buying AI services should therefore ask how much power each task uses and how much all its work needs in total. Together, these figures can show whether gains on one task are helping the business meet its goals for overall demand.
One figure may improve while the other gets worse, with neither report being dishonest. If a business chooses to expand a useful service, it should explain that choice. Reporting only the lower power use per task makes the growth harder to judge. Reporting only the total can hide a gain worth keeping.
Benefits beyond the computing system deserve the same discipline. The IEA's September 21 report on grid modernization describes ways AI and other digital tools can help track and run the grid. They can also help make better use of the network we have. Those are serious reasons to build useful tools. The gains also depend on how and where the tools are used.
A potential grid benefit is not a receipt canceling every other use of AI. It needs evidence about the tool and what it changes. Equally, a power cost does not prove a tool lacks value. We need to weigh both costs and benefits, including what we do not yet know.
That leaves plenty of work for efficiency research. Better hardware, software and ways of working can make a desired level of service less demanding. Planners should account for those gains and change their plans when the evidence changes. They should also keep track of total demand and the supply needed to meet it.
The engineer who halves the electricity required for a task has solved an important problem. If the business then triples the workload, another problem remains. The improvement deserves its credit, and the total still needs to be read.
